Backend Engineer · Production AI Agent Systems
📍 Munich, Germany 🇩🇪  ·  EU Blue Card · authorized to work across the EU

> I build backend systems that run AI agents in production. Specialist
  single agents, supervisor-pattern multi-agent systems, and coding agents
  that ship real artifacts. And I keep them running once customers depend
  on them.

> Team Lead, Backend & GenAI @ innoscripta SE (Munich) — promoted from
  Senior Software Engineer in 9 months. 6+ years backend, 2+ years
  shipping agents that European enterprise customers use daily.

> The hard part was never making an agent work once. It's making it work
  on the ten-thousandth call, with an audit trail, in a regulated domain.

Agents in production, used daily by European enterprise customers.

Architectures

• Specialist single agents — one job, narrow scope, reliable enough to run unattended.
• Multi-agent with a supervisor — a supervisor delegating to specialists, with routing and human approval gates where the blast radius demands one.
• Multi-agent on a custom harness — when a framework's control flow isn't the control flow you need, the harness becomes the product.

Coding agents that deliver artifacts

• Built on LangGraph, DeepAgents, the Claude Agent SDK and OpenHands.
• They produce real deliverables — working code and artifacts a user actually asked for — not a transcript describing what they would do.
• Interfaces generated in real time: the agent builds the UI a user needs while they're using it.

What keeps them alive in production

• MCP servers exposing internal systems as well-typed, observable tools.
• Eval & observability — full-loop traces and LLM-as-judge regression suites, so quality drops are caught by us and not by a customer.
• Memory & context engineering — per-user long-term memory, dynamic context assembly, token budgets.
• Compliance by construction — EU AI Act and GoBD: audit logs, prompt/model versioning, human oversight.

Agents & LLMs

Eval & Observability

Backend & Infra

_{Also shipped with: Kafka · RabbitMQ · Qdrant / Pinecone / Chroma · Elasticsearch · MongoDB · Laravel / PHP · TypeScript / React · Playwright / Scrapy. Plus API security & hardening (auth design, rate limiting, abuse prevention), and reverse engineering as personal research.}

A hands-free voice assistant on an ESP32-S3 with 8 MB of RAM: custom MicroPython firmware, a Claude tool loop with 25 tools, long-term memory in SQLite with embeddings, and a dashboard that shows which tool ran and where every millisecond of the turn went. 751 tests, none of which need the hardware.

The same problems as production agents, on hardware that forgives nothing. Transcription starts before you stop talking, because latency on a voice device is the product and not a metric. Every tool call reports its own provenance instead of collapsing into one opaque "using tools" state. And the device keeps its promises when the brain dies: set a timer, kill the server, it still fires.

▸ github.com/elsayed85/aria-esp32-llm-assistant

🎙️ Build Your Own MCP Server — live on Tek Podcast بالعربي by Ahmed Elemam (53.8K subs) · June 2025 · 7K+ views Built an MCP server from scratch, then a LangGraph agent consuming its tools through an MCP client, with per-step logs. ▸ Watch  ·  Code

innoscripta SE builds Clusterix, a Munich enterprise platform automating R&D documentation, funding applications and audit-ready compliance for automotive, aerospace, biotech, pharma and manufacturing. I lead backend and GenAI engineering for the agent platform.

📧 el.sid.kamal.1999@gmail.com  ·  🔗 linkedin.com/in/elsayed85

_{B.Sc. Computer Science, Mansoura University (2021, Grade A)  ·  Apache Kafka Series, Udemy (2024)  ·  🇪🇬 Arabic native · 🇬🇧 English professional · 🇩🇪 Deutsch elementary}

_{~ Elsayed Abdulrahman}